A data processing method and device, computer equipment and a storage medium

Abstract

Embodiments of the present application disclose a data processing method and device, computer equipment and a storage medium, which can be applied to artificial intelligence and cloud technology scenarios, comprising: obtaining object platform features of a business object; obtaining business model parameters associated with N object groups and object group coding features associated with the N object groups based on a first business model; the first business model and a second business model are obtained after federated learning of sample objects; performing self-attention feature extraction processing on the business model parameters, the object group coding features and the object platform features in the first business model to obtain object group attention features of the business object for the N object groups; and predicting a business index parameter of the business object for candidate business data based on candidate data features of the candidate business data, the object platform features and the object group attention features in the first business model. The embodiments of the present application can improve the data security of the business object.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] In advertising scenarios, a first business model deployed on a first device (e.g., the advertiser's side) and a second business model deployed on a second device (e.g., the advertising platform's side) can be trained using federated learning. When predicting business metrics (e.g., conversion rate) for a business object relative to candidate business data (e.g., a specific ad data), the real-time request sent by the second device to the first device for the business object includes the object identifier of the business object. This allows the first device to predict the business metrics for the business object relative to the candidate ad data based on the object identifier. Since the first and second devices need to transmit the object identifier of the business object during online model application, there is a risk that the object identifier of the business object may be leaked, thus reducing the data security of the business object. Summary of the Invention

[0003] This application provides a data processing method, apparatus, computer equipment, and storage medium, which can improve the data security of business objects.

[0004] One embodiment of this application provides a data processing method, which is executed by a first device and includes:

[0005] Obtain the object platform characteristics of the business object; the object platform characteristics are obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model.

[0006] Based on the first business model, obtain the business model parameters associated with N object groups and the object group encoding features associated with N object groups; N is a positive integer greater than 1; the first business model and the second business model are obtained after federated learning of the sample objects;

[0007] In the first business model, self-attention feature extraction processing is performed on the business model parameters, object group encoding features and object platform features to obtain the object group attention features of the business object for N object groups.

[0008] In the first business model, based on the candidate data features, object platform features, and object group attention features of the candidate business data, the business indicator parameters of the business object for the candidate business data are predicted.

[0009] One embodiment of this application provides a data processing method, which is executed by a first device and includes:

[0010] Based on the first sample information of the sample object on the first device and the first initial model, obtain the sample group coding features associated with N sample groups and the initial model parameters associated with N sample groups; N is the number of object groups configured in the first initial model, and N is a positive integer greater than 1; the sample label of the first sample information is used to characterize the actual indicator parameters of the sample object for the sample business data.

[0011] In the first initial model, self-attention feature extraction is performed on the sample platform features, sample group encoding features, and initial model parameters of the sample object to obtain the sample group attention features of the sample object for N sample groups; the sample platform features are obtained by the second device based on the second initial model and after extracting features from the second sample information of the sample object on the second device.

[0012] In the first initial model, based on the sample data features, sample platform features, and sample group attention features of the sample business data, the prediction index parameters of the sample object for the sample business data are predicted.

[0013] Based on the predicted index parameters, actual index parameters, sample group coding features, and sample group attention features, the total model loss corresponding to the first initial model is determined; the total model loss is used to instruct the second device to perform federated learning training on the second initial model to obtain the second business model.

[0014] Based on the total model loss and the attention features of the sample group, the first initial model is trained by federated learning to obtain the first business model; the first business model is used to predict the business indicator parameters of the business object.

[0015] One embodiment of this application provides a data processing apparatus, including:

[0016] The platform feature acquisition module is used to acquire the object platform features of the business object; the object platform features are obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model.

[0017] The business parameter acquisition module is used to acquire business model parameters associated with N object groups and object group encoding features associated with N object groups based on the first business model; N is a positive integer greater than 1; the first business model and the second business model are obtained by federated learning of sample objects;

[0018] The object group feature determination module is used to perform self-attention feature extraction processing on business model parameters, object group coding features and object platform features in the first business model to obtain the object group attention features of the business object for N object groups.

[0019] The indicator parameter prediction module is used in the first business model to predict the business indicator parameters of the business object for the candidate business data based on the candidate data features, object platform features, and object group attention features of the candidate business data.

[0020] Among them, the object platform feature is obtained by the first device when it receives a traffic request for a business object; the traffic request for the business object belongs to X traffic requests sent by the second device; X is a positive integer; the X traffic requests are selected by the second device based on the initial platform features corresponding to the Y initial objects and the third business model deployed on the second device when generating traffic requests corresponding to Y initial objects respectively; Y is a positive integer greater than or equal to X; the initial platform feature of an initial object is obtained by the second device after extracting features from the object information of an initial object on the second device based on the second business model.

[0021] The business model parameters include object group pooling vectors corresponding to N object groups; the object group encoding features include object group encoding vectors corresponding to N object groups; the N object groups include object group i; i is a positive integer less than or equal to N;

[0022] The object group feature determination module includes:

[0023] The platform feature input unit is used to input the object platform features into the self-attention layer in the first business model;

[0024] The vector acquisition unit is used to obtain the object group pooling vector K of object group i from N object group pooling vectors. i Obtain the object group encoding vector V of object group i from N object group encoding vectors. i ;

[0025] The correlation processing unit is used to process the object platform features and the object group pooling vector K through a self-attention layer. i and object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to object group i is obtained. i ;

[0026] The summation processing unit is used to sum the N attention coefficients through the self-attention layer when the attention coefficients corresponding to each of the N object groups are obtained, so as to obtain the object group attention features of the business object for the N object groups.

[0027] The indicator parameter prediction module includes:

[0028] The splicing processing unit is used to splice the object platform features and object group attention features through the feature splicing layer in the first business model to obtain the business splicing features corresponding to the business object.

[0029] The candidate data feature acquisition unit is used to acquire the candidate data features obtained after feature extraction of the candidate business data.

[0030] The indicator parameter prediction unit is used to input the business splicing features and candidate data features into the multilayer perceptron in the first business model. The multilayer perceptron performs feature extraction processing on the business splicing features and candidate data features to predict the business indicator parameters of the business object for the candidate business data.

[0031] The number of candidate business data is M, where M is a positive integer;

[0032] The device also includes:

[0033] The decision parameter determination module is used to take the M business indicator parameters as M decision parameters when the business object obtains the business indicator parameters for each of the M candidate business data based on the first business model.

[0034] The decision parameter return module is used to return M decision parameters to the second device, so that the second device can select the service data to be sent to the service object from the M candidate service data. The service data to be displayed are the top P candidate service data obtained based on the data sorting result. The data sorting result is obtained by the second device sorting the M candidate service data based on the sorting parameters corresponding to each of the M candidate service data. The sorting parameter of a candidate service data is determined based on the decision parameter of a candidate service data. P is a positive integer less than or equal to M.

[0035] One embodiment of this application provides a data processing apparatus, including:

[0036] The initial parameter acquisition module is used to acquire sample group coding features associated with N sample groups and initial model parameters associated with N sample groups based on the first sample information and the first initial model of the sample objects on the first device; N is the number of object groups configured in the first initial model, and N is a positive integer greater than 1; the sample label of the first sample information is used to characterize the actual indicator parameters of the sample objects for the sample business data;

[0037] The sample group feature determination module is used to perform self-attention feature extraction processing on the sample platform features, sample group encoding features and initial model parameters of the sample object in the first initial model to obtain the sample group attention features of the sample object for N sample groups; the sample platform features are obtained by the second device based on the second initial model after extracting features from the second sample information of the sample object on the second device.

[0038] The prediction module is used to predict the prediction index parameters of the sample objects for the sample business data based on the sample data features, sample platform features, and sample group attention features of the sample business data in the first initial model.

[0039] The loss determination module is used to determine the total model loss corresponding to the first initial model based on the predicted index parameters, actual index parameters, sample group coding features, and sample group attention features; the total model loss is used to instruct the second device to perform federated learning training on the second initial model to obtain the second business model.

[0040] The training module is used to perform federated learning training on the first initial model based on the total model loss and the attention features of the sample group to obtain the first business model; the first business model is used to predict the business indicator parameters of the business object.

[0041] The device also includes:

[0042] The second identifier set acquisition module is used to acquire the second object identifier set associated with the second device based on the sample alignment request when the sample alignment request sent by the second device is received; the object identifiers in the second object identifier set are the object identifiers of historical objects recorded on the second device.

[0043] The first identifier set acquisition module is used to acquire the first object identifier set associated with the first device based on the object identifiers of historical objects recorded on the first device;

[0044] The alignment processing module is used to align the first object identifier set and the second object identifier set to obtain an identifier intersection set. The historical objects corresponding to the identifier objects in the identifier intersection set are used as sample objects that have an intersection relationship with the second device. All identifier objects in the identifier intersection set belong to the first object identifier set and the second object identifier set.

[0045] The identifier intersection return module is used to return the identifier intersection set to the second device so that the second device can determine the sample platform characteristics of the sample objects in the identifier intersection set.

[0046] The second identifier set acquisition module includes:

[0047] The encrypted information acquisition unit is used to acquire the identifier encrypted information in the sample alignment request when the sample alignment request sent by the second device is received; the identifier encrypted information is obtained by the second device encrypting the identifier signature information and the platform object identifier set based on the first device's first public key; the identifier signature information is obtained by the second device signing the platform object identifier set based on the second device's second private key.

[0048] The decryption processing unit is used to decrypt the encrypted identifier information based on the first private key of the first device to obtain the identifier signature information and the platform object identifier set.

[0049] The signature verification processing unit is used to obtain the second public key of the second device, and perform signature verification processing on the identifier signature information based on the second public key to obtain the signature verification result.

[0050] The identifier set determination unit is used to use the platform object identifier set as the second object identifier set associated with the second device when the verification result indicates that the verification is successful.

[0051] The number of sample objects is H, where H is a positive integer greater than or equal to N; the first sample information includes the first sample sub-information of each of the H sample objects on the first device;

[0052] The initial parameter acquisition module includes:

[0053] The grouping unit is used to obtain the number of object groups configured for the first initial model, and to group the H sample objects based on the number of object groups to obtain N sample groups; the sample group includes sample group j; j is a positive integer less than or equal to N;

[0054] The sample group encoding vector determination unit is used to extract the first sample sub-information associated with each sample object in sample group j from H first sample sub-information pieces, and perform feature extraction processing on the obtained first sample sub-information to obtain the sample group encoding vector V corresponding to sample group j. j ;

[0055] The sample group coding feature determination unit is used to obtain the sample group coding features associated with the N sample groups when the sample group coding vectors corresponding to each of the N sample groups are obtained.

[0056] The initial parameter determination unit is used to determine the initial model parameters associated with N sample groups based on the encoded features of the sample groups.

[0057] The loss determination module includes:

[0058] The first loss determination unit is used to determine the first model loss of the first initial model based on the sample group encoding features and the sample group attention features.

[0059] The second loss determination unit is used to determine the second model loss of the second initial model based on the predicted index parameters and the actual index parameters.

[0060] The total loss determination unit is used to superimpose the first model loss and the second model loss to obtain the total model loss corresponding to the first initial model.

[0061] The training module includes:

[0062] The training result determination unit is used to perform federated learning training on the first initial model based on the total model loss to obtain the model training result.

[0063] The parameter adjustment unit is used to adjust the parameters of the initial model based on the attention features of the sample group and the total loss of the model if the model training results indicate that the first initial model after training does not meet the model convergence condition associated with the first initial model.

[0064] The business model determination unit is used to take the adjusted first initial model as a transition model, perform federated learning training on the transition model, and when the transition model after federated learning training meets the model convergence condition, the transition model that meets the model convergence condition is taken as the first business model.

[0065] This application provides a computer device, including: a processor, a memory, and a network interface;

[0066] The processor is connected to a memory and a network interface. The network interface is used to provide data communication functions, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the methods provided in the embodiments of this application.

[0067] One aspect of this application provides a computer-readable storage medium storing a computer program adapted to be loaded and executed by a processor, so that a computer device having the processor performs the method provided in this application.

[0068] One embodiment of this application provides a computer program product, which includes a computer program stored in a computer-readable storage medium; a processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the method of this application embodiment.

[0069] In this embodiment, when predicting the business indicator parameters of a business object for candidate business data, the first device does not need to obtain the object identifier of the business object, but instead obtains the object platform features of the business object sent by the second device. Here, the object platform features are obtained by the second device directly extracting features from the object information of the business object on the second device based on the second device's second business model. Since the first business model trained by federated learning has object grouping capabilities, i.e., it can divide the sample objects participating in training into N object groups, where N is a positive integer greater than 1, the first device can obtain the business model parameters associated with the N object groups and the object group encoding features associated with the N object groups based on the first business model. This allows the first device to perform self-attention feature extraction processing on the business model parameters, object group encoding features, and object platform features in the first business model, thereby obtaining the object group attention features of the business object for the N object groups. Since these object group attention features can be used to replace the object identifier of the business object, this means that when predicting business indicator parameters based on the candidate data features, object platform features, and object group attention features of the candidate business data in the first business model, the risk of leakage of the object identifier of the business object can be greatly reduced, thus improving the data security of the business object. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application;

[0072] Figure 2 This is a schematic diagram illustrating a federated learning-based model application provided in an embodiment of this application;

[0073] Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0074] Figure 4 This is a data interaction diagram in an advertising delivery scenario provided in an embodiment of this application;

[0075] Figure 5 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0076] Figure 6This is a schematic diagram of a scenario for determining a sample object provided in an embodiment of this application;

[0077] Figure 7 This is a structural diagram of a self-attention mechanism provided in an embodiment of this application;

[0078] Figure 8 This is a schematic diagram of model training based on federated learning provided in an embodiment of this application;

[0079] Figure 9 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0080] Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0081] Figure 11 This is a schematic diagram of a computer device provided in an embodiment of this application;

[0082] Figure 12 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application. Detailed Implementation

[0083] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0084] It should be understood that this application provides a federated learning method based on object grouping, which can be applied to the field of artificial intelligence. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or computer-controlled computing to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. Artificial intelligence studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0085] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0086] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0087] Federated learning is a distributed machine learning paradigm that involves two or more participants who conduct joint machine learning through secure algorithmic protocols. It enables joint modeling and inference services from multiple data sources without requiring the original data to be stored or transmitted locally. For example, in this embodiment, the first business model deployed on the first device and the second business model deployed on the second device can be obtained by performing federated learning on sample objects. These sample objects are historical objects recorded in both the first and second devices that have an overlap; that is, the sample object is recorded in both devices.

[0088] Please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. Figure 1 As shown, the network architecture may include server 12F, servers belonging to the same federated platform as server 12F, and a cluster of terminal devices. It is understood that the number of servers belonging to the same federated platform as server 12F can be one or more. For ease of explanation, this embodiment uses one server (e.g., Figure 1 Taking server 11F as an example, this describes a specific implementation method for federated learning between server 11F and server 12F. Figure 1The terminal device cluster shown may include one or more terminal devices, and the terminal device cluster may specifically include terminal device 100a, terminal device 100b, terminal device 100c, ..., terminal device 100n.

[0089] like Figure 1 As shown, terminal devices 100a, 100b, 100c, ..., 100n can each connect to the aforementioned server 11F via a network, enabling each terminal device to interact with the server 11F through this network connection. Furthermore, terminal devices 100a, 100b, 100c, ..., 100n can also connect to the aforementioned server 12F via a network, enabling each terminal device to interact with the server 12F through this network connection. It is understood that the network connection method is not limited; it can be a direct or indirect connection via wired communication, a direct or indirect connection via wireless communication, or other methods. This application does not impose any restrictions on this method.

[0090] Each terminal device in this terminal device cluster can include: smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, in-vehicle terminals, smart TVs, and other smart terminals with data processing capabilities. It should be understood that, for example... Figure 1 Each terminal device in the terminal device cluster shown can have an application client installed. When the application client runs on each terminal device, it can interact with the aforementioned... Figure 1 Data interaction occurs between the servers shown (e.g., server 11F or server 12F). The application client may include social clients, multimedia clients (e.g., video clients), entertainment clients (e.g., game clients), information stream clients, educational clients, live streaming clients, etc. This application client can be a standalone client or an embedded sub-client integrated into another client (e.g., social client, educational client, and multimedia client, etc.), and is not limited thereto.

[0091] like Figure 1 As shown, in this embodiment, servers 11F and 12F can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing cloud computing services. This embodiment does not limit the number of terminal devices and servers.

[0092] It is understood that server 11F and server 12F can each record different object information of historical objects. In the specific embodiments of this application, data related to object information is involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data shall comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0093] In data recommendation scenarios (e.g., advertising scenarios), server 11F can be the computer device corresponding to the advertiser's side, mainly responsible for part of the training of the federated model (e.g., training of the first business model), part of online inference, privacy compliance review, and display of related reports. Advertisers are the initiators of advertising campaigns, merchants who sell or promote their products and services on the internet, and providers of affiliate marketing advertisements. In other words, any merchant promoting or selling their products or services can act as an advertiser, also referred to as the client side in the advertising system. The object information of historical objects recorded by server 11F (i.e., the first object information) can include the object identifier of the historical object (e.g., object ID) and the first historical data information. The first historical data information here can include historical object data information and historical business data information. For example, the historical object data information refers to the goods purchased by the historical object from the advertiser, more detailed characteristics of the goods (goods price, goods reviews, etc.), and tags indicating whether the historical object has converted (e.g., whether the historical object made a purchase or added items to its shopping cart). The historical business data information here can include data information associated with advertising data (e.g., advertising images, advertising titles, etc.).

[0094] In data recommendation scenarios (e.g., advertising scenarios), server 12F can be the computer device corresponding to the advertising platform, mainly responsible for part of the training of the federated model (e.g., training of the second business model), real-time request access, partial online inference, request features, and the distribution of object identifiers for sample groups. The advertising platform refers to a platform used to meet advertisers' needs for advertising and provide traffic distribution. The object information of historical objects recorded by server 12F (i.e., the second object information) can include the object identifier history of historical objects and second historical data information. The second historical data information can include basic information of historical objects and data information shared with server 11F (i.e., the aforementioned historical business data information). The basic information of historical objects can include the age, gender, and contextual characteristics of the historical object (e.g., which articles the historical object viewed before and after triggering the sample business data, the trigger time of the sample business data, and the geographical location of the historical object when triggering the sample business data, etc.).

[0095] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1Among the multiple terminal devices shown, one terminal device is selected as the object device for the business object (e.g., a user accessing the application client in real time). For example, embodiments of this application can... Figure 1 The terminal device 100a shown is an object device. This object device can integrate an application client. This object device can achieve data interaction with the server 11F through the business data platform corresponding to the application client.

[0096] It should be understood that, in the embodiments of this application, the computer device on which the first business model is deployed can be referred to as the first device (e.g., Figure 1 The server 11F shown in the diagram refers to a business model used to predict business metric parameters (e.g., conversion rate, trigger rate, and bid). In this embodiment, the computer device deployed with the second business model can also be referred to as the second device (e.g., Figure 1 The server 12F shown here refers to the second business model, which is used to extract features from object information on the second device. The first and second business models can be obtained by performing federated learning on sample objects. Furthermore, to reduce bandwidth consumption, the second device in this embodiment can also be deployed with a third business model, which is used to reduce the request volume when the number of traffic requests is large (also known as a crowd filtering model).

[0097] When predicting the business indicator parameters of a business object (e.g., object A) relative to candidate business data (e.g., a certain advertisement data), it is necessary to obtain the object platform features of the business object. Here, the object platform features are obtained by server 12F after extracting features from the object information of the business object on server 12F based on the second business model. It can be understood that since the first business model trained by federated learning has an object grouping function—that is, dividing the sample objects participating in training into N object groups, where N is a positive integer greater than 1—server 11F can obtain the business model parameters associated with the N object groups and the object group encoding features associated with the N object groups based on the first business model. Furthermore, server 11F can perform self-attention feature extraction processing on the business model parameters, object group encoding features, and object platform features in the first business model to obtain the object group attention features of the business object relative to the N object groups. Then, based on the candidate data features, object platform features, and object group attention features of the candidate business data, the business indicator parameters of the business object relative to the candidate business data are predicted in the first business model.

[0098] Therefore, in this embodiment, server 11F does not need to obtain the object identifier of the business object. Instead, it obtains the object platform characteristics of the business object through server 12F, so that server 11F can predict business indicator parameters based on the object platform characteristics and the first business model. This can greatly reduce the risk of leakage of the object information of the business object, thereby improving the data security of the business object.

[0099] For better understanding, please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram illustrating a federated learning-based model application provided in an embodiment of this application. For example... Figure 2 As shown, in the advertising scenario, the device 21F in this embodiment can be the first device corresponding to the advertiser, that is, the computer device that deploys the business model 210m (i.e., the first business model). For example, the device 21F can be the one described above. Figure 1 The server 11F shown. In this embodiment, device 22F can be a second device corresponding to the advertising platform side, i.e., a computer device deploying the business model 220m (i.e., the second business model). For example, device 22F can be the one described above. Figure 1 The server shown is 12F.

[0100] Here, business model 210m and business model 220m are obtained after performing federated learning on the sample objects. Figure 2 As shown, the business model 210m may include network layer L1 (Attention, i.e., self-attention layer), network layer L2 (concatenate, i.e., feature concatenation layer), and network layer L3 (Multi-Layer Perception, or MLP for short). The business model 220m may also include a multi-layer perceptron. It should be understood that when device 21F completes federated learning training of the business model 210m, device 21F may refer to the N sample groups into which the sample objects are divided as N object groups, where N is a positive integer greater than 1. For example, these N object groups may specifically include object group 1, object group 2, ..., and object group N.

[0101] Figure 2 The database 2100K shown can be the database corresponding to device 21F. This database 2100K can store the first object information of historical objects. For example, the first object information may include the object identifier of the historical object and the first historical data information of the historical object. Furthermore, when the first business model training is completed, this database 2100K can store business model parameters (object group pooling features, for example) associated with N object groups. Figure 2 The features shown are 2K) and the object group encoding features associated with N object groups (e.g., Figure 2 Feature 2K is shown. This feature 2K can be composed of the object group pooling vector K1 corresponding to object group 1, the object group pooling vector K2 corresponding to object group 2, ..., and the object group pooling vector K corresponding to object group N. N This feature 2V can be composed of the object group encoding vector V1 corresponding to object group 1, the object group encoding vector V2 corresponding to object group 2, ..., and the object group encoding vector V1 corresponding to object group N. N What constitutes it. Figure 2 The database 2200K shown can be the database corresponding to device 22F. This database 2200K can store the second object information of historical objects. For example, the second object information here can include the object identifier of the historical object and the second historical data information of the historical object.

[0102] In an advertising scenario, when a business object accesses an application client associated with the advertising platform corresponding to device 22F, device 21F can obtain the object platform characteristics of the business object sent by device 22F. These object platform characteristics are determined by device 22F based on the object information (e.g., object information 22X) of the business object obtained from database 2200K. Figure 2 The business model 220m shown is obtained after feature extraction of the object information 22X. This object information 22X can include object information in multiple dimensions, such as the object identifier of the business object, the age, gender, contextual features, and historical advertising data information pushed to the business object.

[0103] It should be understood that in the specific embodiments of this application, data related to object information of business objects are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the user (i.e., the business object) is required, and the collection, use and processing of related data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0104] Furthermore, the device 21F can obtain features 2K associated with N object groups and features 2V associated with N object groups from the database 2100K based on the business model 210m. Then, within the business model 210m, it can further analyze features 2K, features 2V, and object platform features (e.g., ...). Figure 2 The feature 2Q shown is subjected to self-attention feature extraction processing to obtain the object group attention feature of the business object for these N object groups (e.g., Figure 2 The feature shown is 2W).

[0105] At the same time, the device 21F can also perform feature extraction processing on the service data 20S to obtain the candidate data features corresponding to the service data 20S (e.g., Figure 2(As shown in feature 2S). At this time, the device 21F can predict the business indicator parameters of the business object for the business data 20S based on feature 2S, the object platform feature, and feature 2W in the business model 210m. In other words, the device 21F can perform feature concatenation processing on feature 2W and the object platform feature in the network layer L2 of the business model 210m to obtain the business concatenation feature (e.g., feature 2P) corresponding to the business object, and then perform feature extraction processing on feature 2P and feature 2S in the network layer L3 of the business model 210m to predict the business indicator parameters of the business object for the business data 20S.

[0106] Therefore, when predicting the business indicator parameters of a business object in relation to business data 20S, it is not necessary to transmit the object identifier of the business object between devices 21F and 22F. Instead, the object platform features obtained by the business model 220m deployed on device 22F are transmitted between the two devices. The object group attention features, which are used to replace the object identifier, are obtained by using the business model 210m deployed on device 21F and the object platform features. This will greatly reduce the risk of leakage of the object information of the business object when predicting the business indicator parameters of the business object using the object group attention features, object platform features, and features 2S of business data 20S, thereby effectively protecting the user privacy of the business object and improving the data security of the business object.

[0107] In this context, the first device deploying the first business model and the second device deploying the second business model, based on federated learning, further employ object grouping to utilize user group information (e.g., object group attention features) instead of the object identifiers of business objects. The specific implementation method for predicting business indicator parameters of business objects for candidate business data can be found below. Figures 3-8 The corresponding implementation examples.

[0108] Further, please see Figure 3 , Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 3 As shown, this method can be executed by a first device, which can be a terminal device deployed with the first business model, or a server deployed with the first business model (e.g., the one described above). Figure 1 The server shown (11F) is not limited here. For ease of understanding, this application embodiment uses the execution of the method by a server with a first business model deployed as an example for illustration. The method may include at least the following steps S101-S104:

[0109] Step S101: Obtain the object platform characteristics of the business object.

[0110] Here, the object platform feature is obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model. Specifically, when a business object accesses an application client associated with the business platform (e.g., an advertising platform) corresponding to the second device, the second device can obtain the object information (i.e., the second object information) of the business object recorded on the second device. Further, the second device can perform feature extraction processing on the object information of the business object based on the second business model deployed on the second device, thereby obtaining the object platform feature of the business object. Then, based on the object platform feature of the business object, a traffic request for the business object can be generated. At this time, the second device can send the traffic request for the business object to the first device, so that the first device can obtain the object platform feature of the business object based on the traffic request.

[0111] Here, the object platform feature is obtained by the first device when it receives a traffic request for a business object; the traffic request for the business object belongs to X traffic requests sent by the second device; X is a positive integer; the X traffic requests are selected by the second device when generating Y initial objects, based on the initial platform features corresponding to the Y initial objects and the third business model deployed on the second device; Y is a positive integer greater than or equal to X. The initial platform feature of an initial object is obtained by the second device after extracting features from the object information of an initial object on the second device based on the second business model.

[0112] It is understood that, at a certain moment, the second device in this application embodiment can generate a large number of traffic requests for multiple real-time business objects accessed on the business platform, and then send a large number of traffic requests to the first device so that the first device can predict the business indicator parameters of each business object for the candidate business data. Of course, since the computing resources of the first device are limited, it cannot process a large number of traffic requests simultaneously. Therefore, in order to reduce the consumption of transmission bandwidth, this application embodiment can deploy a third business model on the second device, namely a crowd filtering model for filtering target groups, thereby reducing the request volume when the number of traffic requests is large.

[0113] For example, if the number of initial objects accessing the business platform at a certain moment is 1000, the second device can extract features from the object information recorded on the second device for each of these 1000 initial objects based on the second business model. The second business model then outputs the initial platform features corresponding to each of the 1000 initial objects. At this point, the second device can generate traffic requests for each initial object based on the initial platform features corresponding to each initial object, thus obtaining 1000 traffic requests.

[0114] Furthermore, the second device can compare the total number of current traffic requests with the number of traffic requests that the first device can process. If the total number of current traffic requests is greater than the processing capacity, the second device can filter the current traffic requests based on the third business model. Optionally, if the total number of current traffic requests is less than or equal to the processing capacity, the second device can directly send all current traffic requests to the first device.

[0115] For example, if the computing resources of the first device can handle 10 traffic requests, this means that even if the second device sends 1000 traffic requests to the first device, the first device cannot process them, resulting in 990 traffic requests being transmitted ineffectively between the two devices. To reduce transmission bandwidth, the second device can input the initial platform characteristics corresponding to the 1000 initial objects into the third business model. The third business model determines the transmission sorting parameters corresponding to these 1000 initial objects, and then sorts the 1000 traffic requests using the transmission sorting parameters of each initial object to obtain the request sorting result. Furthermore, based on the processing capacity of the first device, the second device can select the top 10 traffic requests from these 1000 traffic requests, and then, in an RTA / ADX scenario, send these 10 traffic requests to the first device. Here, RTA (Real Time API) refers to a real-time API, and ADX (ad exchange) refers to an advertising exchange platform. At this point, the first device can arbitrarily select one of the initial object's traffic requests from these 10 traffic requests as the traffic request for the business object, and then obtain the object platform characteristics of the business object from the traffic request of the business object.

[0116] Step S102: Based on the first business model, obtain the business model parameters associated with the N object groups and the object group encoding features associated with the N object groups.

[0117] The first and second business models are obtained by federated learning of sample objects, resulting in N object groups upon training completion. These N object groups can be obtained by the first device grouping sample objects using a self-attention mechanism, or by the first device grouping sample objects using other clustering algorithms; no limitation is imposed here. For ease of understanding, this embodiment can take grouping sample objects based on a self-attention mechanism as an example. Therefore, upon completion of training the first business model, the pooling features of the object groups associated with the N object groups (i.e., the business model parameters of the first business model) and the object group encoding features associated with the N object groups can be stored together. Thus, when the first device obtains the object platform features of the business objects, it can directly and quickly obtain the business model parameters associated with the N object groups (e.g., the aforementioned...) based on the first business model deployed on the first device. Figure 2 The features shown (2K) and the object group encoding features associated with N object groups (e.g., the above) Figure 2 The feature shown is 2V). Here, N is a positive integer greater than 1.

[0118] Step S103: In the first business model, self-attention feature extraction processing is performed on the business model parameters, object group encoding features and object platform features to obtain the object group attention features of the business object for N object groups.

[0119] Here, the business model parameters refer to the object group pooling features associated with N object groups. These object group pooling features can include the object group pooling vectors corresponding to each of the N object groups. The object group encoding features include the object group encoding vectors corresponding to each of the N object groups. The N object groups include object group i, where i is a positive integer less than or equal to N. Specifically, the first device can input the object platform features into the self-attention layer of the first business model, and then obtain the object group pooling vector K of object group i from the N object group pooling vectors. i Obtain the object group encoding vector V of object group i from N object group encoding vectors. i At this point, the first device can use a self-attention layer to analyze the object platform features and the object group pooling vector K. i and object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to object group i is obtained. i When the attention coefficients corresponding to each of the N object groups are obtained, the first device can sum the N attention coefficients through a self-attention layer to obtain the object group attention features of the business object for the N object groups.

[0120] The specific method by which the first device in this application determines the attention features of the object group can be found in the following formulas (1)-(3):

[0121] S i =F(Q,K) i = Similarity(Q,K) i (1)

[0122] A i =softmax(S i (2)

[0123]

[0124] Where Q can be the object platform characteristic of the business object; K i S can be the pooling vector of the object group corresponding to object group i; i This can be used to represent the vector similarity or correlation between the platform features of an object and the pooled vector of object group i. The most common methods include: calculating the dot product of their vectors, calculating their vector cosine similarity, or by introducing an additional neural network. i It can be used to represent the normalization result corresponding to object group i; W i It can be used to represent the attention coefficient corresponding to object group i; Attention can be used to represent the object group attention feature of a business object for N object groups.

[0125] like Figure 2 As shown, feature 2K can be composed of the object group pooling vector K1 corresponding to object group 1, the object group pooling vector K2 corresponding to object group 2, ..., and the object group pooling vector K corresponding to object group N. N It is composed of the object group encoding vector V1 corresponding to object group 1, the object group encoding vector V2 corresponding to object group 2, ..., and the object group encoding vector V1 corresponding to object group N. N What constitutes it.

[0126] When device 21F obtains the object platform characteristics of a business object, it can use the object platform characteristics of that business object (e.g., characteristic 2Q) as input to... Figure 2 The query vector of network layer L1 (i.e., the self-attention layer) is shown. Further, the device 21F can obtain the object group pooling vector K of object group i from the N object group pooling vectors included in feature 2K. i From the N object group encoding vectors included in feature 2V, obtain the object group encoding vector V of object group i. i At this point, device 21F can use network layer L1 to process feature 2Q and object group pooling vector K. iand object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to object group i is obtained. i .

[0127] For example, for object group 1, the device 21F can obtain the object group pooling vector K1 of object group 1 from the N object group pooling vectors included in feature 2K, and obtain the object group encoding vector V1 of object group 1 from the N object group encoding vectors included in feature 2V. Further, in network layer L1, the device 21F can determine the vector similarity between the object platform feature and the object group pooling vector K1 of object group 1 based on the above formula (1), and normalize the vector similarity of object group 1 based on the above formula (2) to obtain the normalized result A1 corresponding to object group 1. Then, based on the above formula (3), the normalized result A1 and the object group encoding vector V1 of object group 1 are multiplied by a dot to obtain the attention coefficient W1 corresponding to object group 1.

[0128] Similarly, the device 21F can obtain the attention coefficients corresponding to each of the N object groups, and then, by summing the N attention coefficients, it can obtain the object group attention features of the business object for the N object groups (e.g., Figure 2 The feature shown is 2W).

[0129] Step S104: In the first business model, based on the candidate data features, object platform features, and object group attention features of the candidate business data, predict the business indicator parameters of the business object for the candidate business data.

[0130] Specifically, the first device can utilize the feature splicing layer in the first business model (e.g., Figure 2 The network layer L2 (as shown) concatenates the object platform features and object group attention features to obtain the business concatenation features corresponding to the business object (e.g., Figure 2 Feature 2P is shown. Simultaneously, the first device can also acquire candidate data features (e.g., feature 2S) obtained after feature extraction of the candidate service data. Further, the first device can input the service splicing features and candidate data features into the multilayer perceptron in the first service model (e.g., Figure 2 The network layer L3 shown uses a multilayer perceptron to extract features from the service splicing features and candidate data features, thereby predicting the service indicator parameters of the service object for the candidate service data.

[0131] It can be understood that the number of candidate business data here can be M, where M is a positive integer. When the first device obtains the business indicator parameters for each of the M candidate business data for the business object based on the first business model, the first device can use the M business indicator parameters as M decision parameters, and then return the M decision parameters to the second device so that the second device can select the business data to be sent to the business object from the M candidate business data. Here, the business data to be displayed is the top P candidate business data obtained by the second device based on the data sorting result; the data sorting result is obtained by the second device sorting the M candidate business data based on the sorting parameters corresponding to each of the M candidate business data; the sorting parameter of a candidate business data is determined based on the decision parameter of a candidate business data; P is a positive integer less than or equal to M. Here, the sorting parameters may include bid, conversion rate, trigger rate, and quality score, etc. In this embodiment, parameters other than decision parameters can be called auxiliary parameters. For example, if the business indicator parameter predicted by the first business model is the conversion rate, the auxiliary parameters may include the trigger rate, bid, and quality score, etc. Among them, conversion rate refers to the probability that a candidate business data will be converted after being clicked in a certain situation by the online advertising system; clickthrough rate refers to the probability that a candidate business data will be clicked after being delivered in a certain situation by the online advertising system; bid refers to the price that an advertiser bids for an ad, that is, the expected cost to obtain a single conversion when the ad is delivered.

[0132] For better understanding, please refer to [link / reference]. Figure 4 , Figure 4 This is a data interaction diagram provided in an advertising placement scenario according to an embodiment of this application. For example... Figure 4 As shown, in the advertising delivery scenario, device 41F in this embodiment can be the first device corresponding to the advertiser, that is, the computer device that deploys the first business model. For example, device 41F can be the aforementioned Figure 1 The server 11F shown. In this embodiment, device 42F can be a second device corresponding to the advertising platform side, i.e., a computer device deploying the second business model. For example, device 42F can be the one described above. Figure 1 The server shown is 12F.

[0133] like Figure 4As shown, the business metric parameters predicted by the first business model deployed in device 41F can be used for advertising placement decisions, including audience targeting, bidding, weighting, and other methods. This first business model refers to the business model associated with the decision-making method. For example, if the decision-making method is audience targeting, then the first business model may include... Figure 4 The model shown is 410m (e.g., Lookalike model, LTR model). The purpose of the Lookalike model is to identify other groups of people similar to the target group from a large population. The LTR model is used for ranking.

[0134] If the decision-making method is a bid-weighted approach, then the first business model may include... Figure 4 The model shown is 420m (e.g., the PKAM model, i.e., the key customer model). The business indicator parameters output by the PKAM model can refer to the weighted conversion rate coefficients for candidate business data.

[0135] If the decision-making method is the first bidding method (e.g., CPA bidding method), then the first business model may include Figure 4 The model 430m shown (e.g., the PLTV model) outputs business metrics parameters that can refer to the lifetime value of users for candidate business data. Here, CPA (Cost Per Action) refers to the cost per action, meaning the pricing method is based on the actual effect of the advertising campaign, i.e., billing based on valid responses to questionnaires or orders, regardless of the amount of advertising.

[0136] If the decision-making method is the second bidding method (e.g., CPC bidding method), then the first business model may include Figure 4 The model shown is 440m (e.g., the PCVR model and the PLTV model). The business metric parameters output by the PCVR model can indicate the conversion rate predicted for candidate business data. Here, CPC (Cost Per Click) refers to the revenue per trigger.

[0137] If the decision-making method is a third bidding method (e.g., CPM bidding method), then the first business model may include Figure 4 The model shown is 450m (e.g., PCTR model, PCVR model, and PLTV model). The PCTR model is used to predict the trigger rate for candidate business data. Here, CPM (cost per mille) refers to revenue per thousand impressions, that is, the advertising revenue that can be obtained for every thousand impressions. The unit of impression can be a webpage, an ad unit, or even a single ad.

[0138] It should be understood that Figure 4In an advertising delivery scenario, device 42F can recall M candidate business data points for a business object, where M is a positive integer. At this time, device 42F can obtain the object platform characteristics of the business object based on the second business model to generate a traffic request for the business object to be sent to device 41F. When device 41F receives the traffic request for the business object, it can obtain the object platform characteristics of the business object to predict the business indicator parameters for the M candidate business data points.

[0139] For example, device 41F can obtain object group attention features for N object groups based on the first business model deployed on device 41F. Then, based on the object group attention features, the candidate data features corresponding to M candidate business data, and the object platform features, it can predict the business indicator parameters of the business object for each candidate business data. Furthermore, device 41F can use these M candidate business data business indicator parameters as M decision parameters to return to device 42F.

[0140] The specific method by which the computer device determines the sorting parameters of the candidate business data can be found in formulas (4)-(5):

[0141] CPM = bid × CTR × CVR (4)

[0142] Ranking_score=CPM+quality (5)

[0143] Among them, CPM refers to the revenue per thousand impressions of candidate business data; bid refers to the bid for candidate business data; CTR refers to the trigger rate of business objects for candidate business data; CVR refers to the conversion rate of business objects for candidate business data; quality refers to the quality score of candidate business data; and Ranking_score refers to the ranking score used to rank candidate business data.

[0144] For example, during the initial selection process, device 42F can directly use the revenue per thousand impressions determined by formula (4) as the ranking parameter for candidate business data. Optionally, during the final selection process, device 42F can use the ranking score obtained based on formulas (4) and (5) as the ranking parameter for candidate business data, without any restrictions. It is understood that the higher the value of the ranking parameter, the higher the ranking of the candidate business data.

[0145] It is understandable that device 42F can select one candidate business data from M candidate business data as the data to be sorted, and then determine the sorting parameters of the data to be sorted based on the business indicator parameters of the data to be sorted. The sorting parameters here can be determined according to the above formulas (4)-(5), based on the bid corresponding to the data to be sorted, the conversion rate of the business object for the data to be sorted, the trigger rate of the business object for the data to be sorted, and the quality score of the candidate business data.

[0146] Upon obtaining the sorting parameters corresponding to M candidate service data, the device 42F can sort the M candidate service data based on the M sorting parameters to obtain a data sorting result. This data sorting result can be used to indicate the display priority order of the M candidate service data; that is, the higher the value of the sorting parameter, the earlier the corresponding candidate service data is displayed. Furthermore, based on the data sorting result, the device 42F can obtain the top P candidate service data and use these P candidate service data as the service data to be sent to the service object for display.

[0147] For example, if the number M of candidate business data is 200, the computer device can obtain the top 10 candidate business data based on the data sorting result after sorting these 200 candidate business data, and use these 10 candidate business data as the business data to be displayed to the business object (i.e., the business data recommended to the business object).

[0148] Therefore, in this embodiment, when device 42F sends a traffic request in an RTA / ADX scenario, it simultaneously uses a model to infer the object platform characteristics of the business object online. This helps the advertiser-side demand-side platform (DSP) of the RTA / ADX, which participates in the advertising ranking competition, to better predict business indicator parameters and improve the return on investment (ROI) of advertising. The goal of joint modeling is to help optimize the first business model of device 41F (i.e., the advertiser side) to improve the accuracy of business indicator prediction and enhance advertising effectiveness while protecting user privacy, thereby promoting advertising consumption. Simultaneously, the decision results of device 41F can be mined as features to help optimize the performance of the various models deployed on device 42F.

[0149] In this embodiment, when predicting the business indicator parameters of a business object for candidate business data, the first device does not need to obtain the object identifier of the business object, but instead obtains the object platform features of the business object sent by the second device. Here, the object platform features are obtained by the second device directly extracting features from the object information of the business object on the second device based on the second device's second business model. Since the first business model trained by federated learning has object grouping capabilities, i.e., it can divide the sample objects participating in training into N object groups, where N is a positive integer greater than 1, the first device can obtain the business model parameters associated with the N object groups and the object group encoding features associated with the N object groups based on the first business model. This allows the first device to perform self-attention feature extraction processing on the business model parameters, object group encoding features, and object platform features in the first business model, thereby obtaining the object group attention features of the business object for the N object groups. Since these object group attention features can be used to replace the object identifier of the business object, this means that when predicting business indicator parameters based on the candidate data features, object platform features, and object group attention features of the candidate business data in the first business model, the risk of leakage of the object identifier of the business object can be greatly reduced, thus improving the data security of the business object.

[0150] Further, please see Figure 5 , Figure 5 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method can be executed by a first device deployed with a first business model; the first device can be one of the aforementioned... Figure 1 Any terminal device in the terminal device cluster shown, for example, terminal device 100a), can execute the above-mentioned... Figure 1 The server 11F shown can execute the algorithm, and it can also be executed interactively by a terminal device with model application capabilities and a server with model training capabilities. No limitation is made here. This method may include at least the following steps S201-S209:

[0151] Step S201: Obtain the object platform characteristics of the business object.

[0152] Here, the object platform feature is obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model. Specifically, when a business object accesses an application client associated with the business platform (e.g., an advertising platform) corresponding to the second device, the second device can obtain the object information (i.e., the second object information) of the business object recorded on the second device. Further, the second device can perform feature extraction processing on the object information of the business object based on the second business model deployed on the second device, thereby obtaining the object platform feature of the business object. Then, based on the object platform feature of the business object, a traffic request for the business object can be generated. At this time, the second device can send the traffic request for the business object to the first device, so that the first device can obtain the object platform feature of the business object based on the traffic request.

[0153] Step S202: Based on the first business model, obtain the business model parameters associated with the N object groups and the object group encoding features associated with the N object groups.

[0154] The first business model and the second business model are obtained by federated learning of the sample objects. It is understood that, since the first device can store the object group pooling features (i.e., the business model parameters of the first business model) associated with N object groups and the object group encoding features associated with N object groups together when training the first business model, the first device can directly and quickly obtain the business model parameters associated with the N object groups (e.g., the aforementioned) based on the first business model deployed on the first device when it obtains the object platform features of the business objects. Figure 2 The features shown (2K) and the object group encoding features associated with N object groups (e.g., the above) Figure 2 The feature shown is 2V). Here, N is a positive integer greater than 1.

[0155] Step S203: In the first business model, self-attention feature extraction processing is performed on the business model parameters, object group encoding features and object platform features to obtain the object group attention features of the business object for N object groups.

[0156] Here, the business model parameters refer to the object group pooling features associated with N object groups. These object group pooling features can include the object group pooling vectors corresponding to each of the N object groups. The object group encoding features include the object group encoding vectors corresponding to each of the N object groups. The N object groups include object group i, where i is a positive integer less than or equal to N. Specifically, the first device can input the object platform features into the self-attention layer of the first business model, and then obtain the object group pooling vector K of object group i from the N object group pooling vectors.i Obtain the object group encoding vector V of object group i from N object group encoding vectors. i At this point, the first device can use a self-attention layer to analyze the object platform features and the object group pooling vector K. i and object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to object group i is obtained. i When the attention coefficients corresponding to each of the N object groups are obtained, the first device can sum the N attention coefficients through a self-attention layer to obtain the object group attention features of the business object for the N object groups.

[0157] Step S204: In the first business model, based on the candidate data features, object platform features, and object group attention features of the candidate business data, predict the business indicator parameters of the business object for the candidate business data.

[0158] Specifically, the first device can utilize the feature splicing layer in the first business model (e.g., Figure 2 The network layer L2 (as shown) concatenates the object platform features and object group attention features to obtain the business concatenation features corresponding to the business object (e.g., Figure 2 Feature 2P is shown. Simultaneously, the first device can also acquire candidate data features (e.g., feature 2S) obtained after feature extraction of the candidate service data. Further, the first device can input the service splicing features and candidate data features into the multilayer perceptron in the first service model (e.g., Figure 2 The network layer L3 shown uses a multilayer perceptron to extract features from the service splicing features and candidate data features, thereby predicting the service indicator parameters of the service object for the candidate service data.

[0159] The data processing method in this embodiment may include a model training process (i.e., the offline stage) and a model application process (i.e., the online stage). The offline stage may include sample intersection, joint training, and compliance review. The online stage may include model deployment and upper bounds, and the establishment of data links for online inference. It is understood that steps S201-S204 describe the model application process, and the specific implementation of this process can be found above. Figure 3 The descriptions of steps S101-S104 in the corresponding embodiments will not be repeated here.

[0160] The specific training process of this model can be found in the description of steps S205-S209 below.

[0161] Step S205: Based on the first sample information of the sample object on the first device and the first initial model, obtain the sample group coding features associated with N sample groups and the initial model parameters associated with N sample groups.

[0162] Here, N refers to the number of object groups configured for the first initial model, and N is a positive integer greater than 1. The sample label of the first sample information can be used to characterize the actual indicator parameters of the sample object for the sample business data. Here, the number of sample objects is H, and H is a positive integer greater than or equal to N; the first sample information includes the first sample sub-information of H sample objects on the first device. Specifically, the first device can obtain the number of object groups configured for the first initial model, and then group the H sample objects based on the number of object groups to obtain N sample groups. Here, the sample group can include sample group j; j is a positive integer less than or equal to N. Further, the first device can obtain the first sample sub-information associated with each sample object in sample group j from the H first sample sub-information, and perform feature extraction processing on the obtained first sample sub-information to obtain the sample group encoding vector V corresponding to sample group j. j When the first device obtains the sample group encoding vector corresponding to each of the N sample groups, it can obtain the sample group encoding features associated with the N sample groups, and then determine the initial model parameters associated with the N sample groups based on the sample group encoding features.

[0163] It should be understood that when the first device and the second device perform federated learning on the first and second initial models on the federated platform, the first device can obtain the sample alignment request sent by the second device, and then obtain the second object identifier set associated with the second device based on the sample alignment request. Here, the object identifiers in the second object identifier set are the object identifiers of historical objects recorded on the second device. To effectively ensure the security of object identifiers during data transmission, in this embodiment, when the second device sends its recorded object identifier set to the first device, it can process the object identifier set using an encryption algorithm before transmission. Optionally, the second device can also encode the object identifier set based on identifier encoding rules before transmission. This is not limited here. The method by which the first device encodes the object identifiers is the same as the method by which the second device encodes the object identifiers.

[0164] For ease of understanding, this application embodiment can be described as processing the object identifier set using an encryption algorithm before transmission. Specifically, when the first device receives a sample alignment request sent by the second device, it can obtain the identifier encryption information in the sample alignment request. This identifier encryption information is obtained by the second device encrypting the identifier signature information and the platform object identifier set using the first device's first public key. The identifier signature information can be obtained by the second device signing the platform object identifier set using its second private key. Further, the first device can decrypt the identifier encryption information using its first private key to obtain the identifier signature information and the platform object identifier set. Then, it can obtain the second device's second public key and perform signature verification processing on the identifier signature information to obtain a verification result. When the verification result indicates successful verification, the first device can use the platform object identifier set as the second object identifier set associated with the second device.

[0165] Simultaneously, the first device can also obtain a first object identifier set associated with the first device based on the object identifiers of historical objects recorded on the first device. Further, the first device can align the first object identifier set and the second object identifier set to obtain an identifier intersection set, using the historical objects corresponding to the identifier objects in the identifier intersection set as sample objects that have an intersection relationship with the second device; wherein, all identifier objects in the identifier intersection set belong to both the first and second object identifier sets. At this point, the first device can return the identifier intersection set to the second device so that the second device can determine the sample platform characteristics of the sample objects in the identifier intersection set.

[0166] For better understanding, please refer to [link / reference]. Figure 6 , Figure 6 This is a schematic diagram illustrating a scenario for determining a sample object, as provided in an embodiment of this application. Figure 6 As shown, in the advertising delivery scenario, device 61F in this embodiment can be the first device corresponding to the advertiser, that is, the computer device that deploys the first business model. For example, device 61F can be the aforementioned Figure 1 The server 11F shown. In this embodiment, device 62F can be a second device corresponding to the advertising platform side, i.e., a computer device deploying the second business model. For example, device 62F can be the one described above. Figure 1 The server shown is 12F.

[0167] like Figure 6As shown, when performing federated learning on the first initial model deployed by device 61F and the second initial model deployed by device 62F on the federated platform, device 62F can obtain the object identifiers of historical objects recorded on device 62F from the database 6200K corresponding to device 62F. The obtained object identifiers can then be used as the set of object identifiers associated with device 62F (i.e., the platform object identifier set). Here, the platform object identifier set can take six items as an example, specifically including object identifier B. a Object identifier B b Object identifier B c Object identifier B d Object identifier B e and object identifier B f .

[0168] To effectively ensure data security when object identifiers are transmitted between two devices, device 62F can sign the platform object identifier set based on its private key (i.e., the second private key) to obtain identifier signature information. It is understood that device 62F can obtain a hash calculation rule for the platform object identifier set, which can be a pre-agreed digest algorithm between device 62F and other devices in the federated platform (e.g., device 61F). Therefore, device 62F can perform a hash calculation on the platform object identifier set based on this hash calculation rule to obtain digest information (e.g., digest information h) of the platform object identifier set. In this embodiment, the digest information of the platform object identifier set determined by device 62F can be referred to as the first digest information. Further, device 62F can sign the first digest information based on its private key to obtain the identifier signature information. Furthermore, the device 62F can also obtain the public key (i.e. the first public key) of the device 61F, encrypt the identifier signature information and the platform object identifier set to obtain the identifier encrypted information, and then generate a sample alignment request to be sent to the device 61F using the identifier encrypted information.

[0169] When device 61F receives the sample alignment request, it can obtain the encrypted identifier information in the request. Then, based on its private key (i.e., the first private key), device 61F can decrypt the encrypted identifier information to obtain the identifier signature information and the platform object identifier set. At this point, device 61F can obtain device 62F's public key (i.e., the second public key) and, based on this second public key, verify the tag signature information to obtain the verification result.

[0170] It is understood that device 61F can verify the signature information based on the public key of device 62F to obtain the first digest information. Simultaneously, device 61F can also obtain the same hash calculation rules as device 62F to perform hash calculation on the platform object identifier set, thereby obtaining the digest information (e.g., digest information H) of the platform object identifier set. In this embodiment, the digest information of the platform object identifier set determined by device 61F can be referred to as the second digest information. At this time, device 61F can compare the first digest information with the second digest information to obtain the verification result, thereby determining whether the platform object identifier set has been tampered with. It is understood that if the first digest information and the second digest information are different, device 61F can determine that the verification result indicates verification failure. Optionally, if the first digest information is the same as the second digest information, the device 61F can determine that the signature verification result indicates successful signature verification. This means that the platform object identifier set has not been tampered with and that the platform object identifier set was indeed sent by the device 62F. In this case, the device 61F can use the platform object identifier set as the second object identifier set associated with the device 62F.

[0171] Simultaneously, the first device can also retrieve the object identifiers of historical objects recorded on device 61F from the database 6100K corresponding to device 61F, and then use the retrieved object identifiers as a set of object identifiers associated with device 61F (i.e., the first object identifier set). Here, the first object identifier set can take five items as an example, specifically including object identifier B. a Object identifier B c Object identifier B d Object identifier B e and object identifier B g .

[0172] At this point, device 61F can align the first object identifier set and the second object identifier set to obtain an identifier intersection set. Specifically, device 61F can add object identifiers that coexist in both the first and second object identifier sets to the identifier intersection set, and then use the historical objects corresponding to the identifier objects in the identifier intersection set as sample objects that have an intersection relationship with the second device. Figure 6 The intersection set of identifiers shown may include object identifier B a Object identifier B c Object identifier B d Object identifier B e That is, the sample object here can include object identifier B. a Corresponding object a, object identifier B c The corresponding object c, object identifier B d The corresponding object d, object identifier Be The corresponding object e.

[0173] Furthermore, device 61F can return the set of identifier intersection to device 62F, so that the second device can input the object information (i.e. the second sample information) recorded on device 62F to the second initial model based on the object identifier of the sample object in the set of identifier intersection. The second initial model outputs the sample platform features corresponding to each sample object, and then sends the sample platform features corresponding to each sample object to device 61F, so that device 61F can perform federated learning training on the first initial model.

[0174] It should be understood that when the first device acquires H sample objects, it can obtain the number of object groups pre-configured for the first initial model. This number of object groups can be the number of clusters obtained by clustering the sample features corresponding to the object information (i.e., the first sample information) on the first device using a clustering algorithm. For example, the number of object groups can be N, where N is a positive integer greater than 1, and H is a positive integer greater than or equal to N.

[0175] Furthermore, the first device can group the H sample objects based on the number of object groups to obtain N sample groups. Grouping refers to the process of anonymizing objects and then aggregating them according to their attributes to calculate their statistical properties. Here, a sample group may include sample group j; j is a positive integer less than or equal to N. Further, the first device can extract first sample sub-information associated with each sample object in sample group j from the H first sample sub-information, and perform feature extraction processing on the obtained first sample sub-information to obtain the sample group encoding vector V corresponding to sample group j. j Upon obtaining the sample group encoding vector corresponding to each of the N sample groups, the first device can obtain the sample group encoding features associated with the N sample groups (i.e., the Value in the self-attention mechanism). Based on these features, it can then determine the initial model parameters associated with the N sample groups (the Key in the self-attention mechanism). For example, the first device can directly perform pooling on the sample group encoding features to obtain the sample group pooling features associated with the N sample groups, which are the initial model parameters of the first initial model.

[0176] Step S206: In the first initial model, self-attention feature extraction processing is performed on the sample platform features, sample group encoding features and initial model parameters of the sample object to obtain the sample group attention features of the sample object for N sample groups.

[0177] Here, the sample platform features are obtained by the second device based on the second initial model, after extracting features from the second sample information of the sample object on the second device. The initial model parameters include sample group pooling vectors corresponding to N sample groups, and the sample group encoding features include sample group encoding vectors corresponding to N sample groups. Specifically, the first device can input the sample platform features into the self-attention layer in the first initial model, and then obtain the sample group pooling vector K of sample group j from the N sample group pooling vectors. j Obtain the sample group encoding vector V of sample group j from N sample group encoding vectors. j At this point, the first device can use the self-attention layer of the first initial model to analyze the sample platform features and the sample group pooling vector K. j and the sample group encoding vector V j After performing correlation processing, the attention coefficient W corresponding to sample group j is obtained. j When the attention coefficients corresponding to each of the N sample groups are obtained, the first device can sum the N attention coefficients through a self-attention layer to obtain the sample group attention features of the sample object for the N sample groups.

[0178] It should be understood that the embodiments of this application can use a self-attention mechanism to replace clustering operations, enabling the model to achieve end-to-end learning, with a single model training iteration fulfilling all functional requirements. For further understanding, please refer to [link to relevant documentation]. Figure 7 , Figure 7 This is a structural diagram of a self-attention mechanism provided in an embodiment of this application. For example... Figure 7 As shown in the embodiments of this application, the constituent elements of the self-attention mechanism can be viewed as being composed of a series of sample groups corresponding to...<K,V> The data pairs are structured as follows. For example, there are four sample groups: sample group 1, sample group 2, sample group 3, and sample group 4. The initial model parameters (i.e., the pooling features associated with the four sample groups) can be composed of the pooling vector 7K1 corresponding to sample group 1, the pooling vector 7K2 corresponding to sample group 2, ..., and the pooling vector K4 corresponding to sample group 4. The encoding features associated with the four sample groups can be composed of the encoding vector 7V1 corresponding to sample group 1, the encoding vector 7V2 corresponding to sample group 2, ..., and the encoding vector V4 corresponding to sample group 4.

[0179] When obtaining the sample platform features of the sample object, the first device can regard the sample platform features as the Query in the self-attention mechanism. Further, the first device obtains the weight coefficient of each sample group by determining the correlation between the sample platform features and the sample group encoding features (Value), that is, the similarity or correlation between the sample platform features and the sample group encoding vectors of each sample group. Then, the initial model parameters (i.e., sample group pooling features) are weighted and summed to obtain the final Attention value (i.e., sample group attention features).

[0180] Step S207: In the first initial model, based on the sample data features, sample platform features, and sample group attention features of the sample business data, predict the prediction index parameters of the sample object for the sample business data.

[0181] Specifically, the first device can concatenate the sample platform features and sample group attention features through the feature concatenation layer in the first initial model to obtain the sample concatenation features corresponding to the sample object. Simultaneously, the first device can also acquire the sample data features obtained after feature extraction of the sample business data. Further, the first device can input the sample concatenation features and sample data features into the multilayer perceptron in the first initial model, and perform feature extraction processing on the sample concatenation features and sample data features through the multilayer perceptron, thereby predicting the prediction index parameters of the sample object for the sample business data.

[0182] Step S208: Based on the predicted index parameters, actual index parameters, sample group coding features, and sample group attention features, determine the total model loss corresponding to the first initial model.

[0183] Here, the total model loss can be used to instruct the second device to perform federated learning training on the second initial model to obtain a second business model. This second business model can be used to predict the object platform features of the business object. The first device can determine the first model loss of the first initial model based on the sample group encoding features and sample group attention features; then, it can determine the second model loss of the second initial model based on the predicted index parameters and the actual index parameters. By superimposing the first model loss and the second model loss, the total model loss corresponding to the first initial model is obtained.

[0184] Step S209: Based on the total model loss and the attention features of the sample group, perform federated learning training on the first initial model to obtain the first business model.

[0185] Specifically, the first device can perform federated learning training on a first initial model based on the total model loss to obtain the model training result. If the model training result indicates that the trained first initial model meets the model convergence condition associated with the first initial model, the first device can directly use the first initial model as the first business model. Optionally, if the model training result indicates that the trained first initial model does not meet the model convergence condition associated with the first initial model, the first device can adjust the parameters of the initial model based on the attention features of the sample group and the total model loss, and then use the adjusted first initial model as a transition model. Federated learning training is then performed on the transition model until the transition model trained by federated learning meets the model convergence condition. The transition model that meets the model convergence condition is then used as the first business model. The first business model is used to predict the business indicator parameters of the business object.

[0186] The specific method for adjusting the model parameters of the first initial model can be found in the following formula (6):

[0187] K t =α*K t-1 +(1-α)*W (6)

[0188] Wherein, K t K refers to the model parameters of the first initial model used in federated learning. t-1 This refers to the model parameters of the initial model during the previous federated learning training; α is a constant within the range of a first threshold to a second threshold, where the first threshold can be 0.7 and the second threshold can be 0.9. These first and second thresholds can be dynamically adjusted according to business needs, and will not be limited here. W refers to the sample group attention features of the sample object for N sample groups.

[0189] For better understanding, please refer to [link / reference]. Figure 8 , Figure 8 This is a schematic diagram illustrating model training based on federated learning, provided in an embodiment of this application. For example... Figure 8 As shown, in the advertising scenario, the device 81F in this embodiment can be the first device corresponding to the advertiser, that is, the computer device that deploys the model 810m (i.e., the first initial model). For example, the device 81F can be the one described above. Figure 1 The server 11F shown. In this embodiment, device 82F can be a second device corresponding to the advertising platform side, i.e., a computer device deploying model 820m (i.e., the second initial model) and model 830m (i.e., the third initial model). For example, device 82F can be the aforementioned... Figure 1 The server shown is 12F.

[0190] Specifically, devices 81F and 82F can perform secure sample alignment based on the federated platform to obtain an identifier intersection set. The historical objects corresponding to the object identifiers in this identifier intersection set can then be used as sample objects for federated learning training. Device 82F can perform feature extraction processing on the second sample information (e.g., sample information 82X) recorded on the sample object based on model 820m to obtain the sample platform features corresponding to the sample object. These sample platform features can then be sent to device 81F.

[0191] like Figure 8 As shown, model 810m may include a self-attention layer, a feature concatenation layer, and a multilayer perceptron. When device 81F acquires sample objects, it can use the first sample information of H sample objects on device 81F (e.g., Figure 8 The sample information 81X shown and the model 810m are used to obtain the sample group encoding features associated with N sample groups (e.g., Figure 8 The features shown are 8V) and the initial model parameters associated with N sample groups (i.e., sample group pooling features, for example, Figure 8 The 8K features shown here can be the features obtained after pooling the encoded features of the sample group.

[0192] Furthermore, the device 81F can perform self-attention feature extraction processing on the sample platform features, feature 8K, and feature 8V of the sample object in the self-attention layer of model 810m to obtain the sample group attention features (e.g., feature 8W) of the sample object for N sample groups. Further, the device 81F can perform feature concatenation processing on the sample platform features and feature 8W through the feature concatenation layer of model 810m to obtain the sample concatenation features (e.g., feature 8P) corresponding to the sample object. Then, feature 8P and the sample data features corresponding to the sample business data (e.g., feature 8S) are input together into the multilayer perceptron in model 810m to predict the prediction index parameters of the sample object for the sample business data.

[0193] It should be understood that the model convergence condition obtained by device 81F refers to the business condition used to indicate that the model stops training. The model convergence condition can be that the number of training times reaches the training time threshold set by the model convergence condition (e.g., 100 times), or it can be that the current model total loss reaches the loss threshold in the model convergence condition. Here, we will not limit it.

[0194] Understandably, the device 81F can determine the first model loss of model 810m based on feature 8W and sample group encoded features (e.g., Figure 8The model loss shown is 10L). Based on the predicted index parameters and the actual index parameters indicated by the sample labels, the second model loss of model 810m is determined (e.g., Figure 8 The model loss 20L is shown, and then the model loss 10L and model loss 20L are superimposed to obtain the total model loss of model 810m. Further, the device 81F can train the first initial model based on the total model loss to obtain the model training result. It is understood that if the model training result indicates that the trained model 810m meets the model convergence condition, the device 81F can directly use the model 810 as the first business model. Optionally, if the model training result indicates that the trained model 810m does not meet the model convergence condition, the device 81F can backpropagate to calculate the gradient based on the sample group attention features and the total model loss, and adjust the model parameters of model 810m based on the gradient and the above formula (6), and then use the adjusted model 810m as a transition model, perform federated learning training on the transition model until the transition model after federated learning training meets the model convergence condition, and then use the transition model that meets the model convergence condition as the first business model.

[0195] Understandably, device 81F can also determine the return parameters to be returned to device 82F based on the total model loss. For example, device 81F can directly use the total model loss as the return parameter, so that device 82F can perform federated learning training on model 820m based on this total model loss. Optionally, device 81F can use the gradient determined based on the total model loss as the return parameter, so that device 82F can directly perform federated learning training on model 820m based on this gradient. The method by which device 82F performs federated learning training on model 820m is similar to the method by which device 81F performs federated learning training on model 810m, and will not be elaborated further here. In addition, device 82F can also train model 830m based on the output results of model 820m (e.g., sample platform features of sample objects) when model 820m is trained, and then train the model parameters of model 830m based on the sample platform features of sample objects and the sample labels of sample objects for model 830m (e.g., whether a traffic request is sent) to obtain a third business model for reducing the request volume.

[0196] In this embodiment, during online inference, the first device and the second device do not need to transmit the object identifier of the business object. Instead, the second device transmits the object platform features determined by the second device based on the second business model. Subsequently, the first device obtains anonymized object group information (i.e., the object group attention features of the business object for N objects) through the object platform features and the first business model. This greatly reduces the risk of leakage of the object identifier of the business object, thereby improving the data security of the business object. In other words, it further protects object privacy while maintaining model accuracy. During federated learning training, by grouping sample objects, N object groups associated with the first business model are obtained upon completion of training. This improves prediction accuracy and efficiency when predicting business indicator parameters of a large number of business objects subsequently.

[0197] Further, please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Figure 9 As shown, the data processing device 1 can be a computer program (including program code) running on a computer device; for example, the data processing device 1 is an application software. The data processing device 1 can be used to execute the corresponding steps in the methods provided in the embodiments of this application. Figure 9 As shown, the data processing device 1 can operate on a first device, which is a computer device deployed with a first business model. The first device can be the aforementioned... Figure 1 The server 11F in the corresponding embodiment can also be the one described above. Figure 1 The corresponding embodiment refers to any terminal device in the terminal device cluster running the first business model, such as terminal device 100a. The data processing device 1 may include: a platform feature acquisition module 10, a business parameter acquisition module 20, an object group feature determination module 30, an indicator parameter prediction module 40, a decision parameter determination module 50, and a decision parameter return module 60.

[0198] The platform feature acquisition module 10 is used to acquire the object platform features of the business object; the object platform features are obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model.

[0199] The business parameter acquisition module 20 is used to acquire business model parameters associated with N object groups and object group encoding features associated with N object groups based on the first business model; N is a positive integer greater than 1; the first business model and the second business model are obtained by federated learning of sample objects;

[0200] The object group feature determination module 30 is used to perform self-attention feature extraction processing on business model parameters, object group encoding features and object platform features in the first business model to obtain the object group attention features of the business object for N object groups.

[0201] The business model parameters include object group pooling vectors corresponding to N object groups; the object group encoding features include object group encoding vectors corresponding to N object groups; the N object groups include object group i; i is a positive integer less than or equal to N;

[0202] The object group feature determination module 30 includes: a platform feature input unit 301, a vector acquisition unit 302, a correlation processing unit 303, and a summation processing unit 304.

[0203] The platform feature input unit 301 is used to input the object platform features into the self-attention layer in the first business model;

[0204] The vector acquisition unit 302 is used to obtain the object group pooling vector K of object group i from N object group pooling vectors. i Obtain the object group encoding vector V of object group i from N object group encoding vectors. i ;

[0205] The correlation processing unit 303 is used to process the object platform features and the object group pooling vector K through a self-attention layer. i and object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to object group i is obtained. i ;

[0206] The summation processing unit 304 is used to sum the N attention coefficients through a self-attention layer when the attention coefficients corresponding to each of the N object groups are obtained, so as to obtain the object group attention features of the business object for the N object groups.

[0207] The specific implementation methods of the feature input unit 301, vector acquisition unit 302, correlation processing unit 303, and summation processing unit 304 of this platform can be found in the above description. Figure 3 The description of step S103 in the corresponding embodiment will not be repeated here.

[0208] The indicator parameter prediction module 40 is used to predict the business indicator parameters of the business object for the candidate business data based on the candidate data features, object platform features and object group attention features of the candidate business data in the first business model.

[0209] The indicator parameter prediction module 40 includes: a splicing processing unit 401, a candidate data feature acquisition unit 402, and an indicator parameter prediction unit 403.

[0210] The splicing processing unit 401 is used to splice the object platform features and object group attention features through the feature splicing layer in the first business model to obtain the business splicing features corresponding to the business object.

[0211] The candidate data feature acquisition unit 402 is used to acquire the candidate data features obtained after feature extraction of the candidate business data;

[0212] The indicator parameter prediction unit 403 is used to input the business splicing features and candidate data features into the multilayer perceptron in the first business model, and to perform feature extraction processing on the business splicing features and candidate data features through the multilayer perceptron to predict the business indicator parameters of the business object for the candidate business data.

[0213] The specific implementation methods of the splicing processing unit 401, the candidate data feature acquisition unit 402, and the indicator parameter prediction unit 403 can be found in the above description. Figure 3 The description of step S104 in the corresponding embodiment will not be repeated here.

[0214] Among them, the object platform feature is obtained by the first device when it receives a traffic request for a business object; the traffic request for the business object belongs to X traffic requests sent by the second device; X is a positive integer; the X traffic requests are selected by the second device based on the initial platform features corresponding to the Y initial objects and the third business model deployed on the second device when generating traffic requests corresponding to Y initial objects respectively; Y is a positive integer greater than or equal to X; the initial platform feature of an initial object is obtained by the second device after extracting features from the object information of an initial object on the second device based on the second business model.

[0215] The number of candidate business data is M, where M is a positive integer;

[0216] The decision parameter determination module 50 is used to take the M business indicator parameters as M decision parameters when the business object obtains the business indicator parameters of each candidate business data in the M candidate business data based on the first business model.

[0217] The decision parameter return module 60 is used to return M decision parameters to the second device, so that the second device can select the service data to be sent to the service object from the M candidate service data; the service data to be displayed is the top P candidate service data obtained based on the data sorting result; the data sorting result is obtained by the second device sorting the M candidate service data based on the sorting parameters corresponding to the M candidate service data respectively; the sorting parameter of a candidate service data is determined based on the decision parameter of a candidate service data; P is a positive integer less than or equal to M.

[0218] The specific implementation methods of the platform's feature acquisition module 10, business parameter acquisition module 20, object group feature determination module 30, indicator parameter prediction module 40, decision parameter determination module 50, and decision parameter return module 60 can be found in the above description. Figure 3 The descriptions of steps S101-S104 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0219] Further, please see Figure 10 , Figure 10 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application. The data processing apparatus 2 can be a computer program (including program code) running on a computer device; for example, the data processing apparatus 2 is application software. The data processing apparatus 2 can be used to execute corresponding steps in the method provided in the embodiments of this application. Figure 10 As shown, the data processing device 2 can operate on a first device, which is a computer device deployed with a first business model. The first device can be the aforementioned... Figure 1 The server 11F in the corresponding embodiment can also be the one described above. Figure 1 Any terminal device running the first business model in the terminal device cluster in the corresponding embodiment, for example, terminal device 100a. The data processing device 2 may include: an initial parameter acquisition module 100, a sample group feature determination module 200, a prediction module 300, a loss determination module 400, a training module 500, a second identifier set acquisition module 600, a first identifier set acquisition module 700, an alignment processing module 800, and an identifier intersection return module 900.

[0220] The initial parameter acquisition module 100 is used to acquire sample group coding features associated with N sample groups and initial model parameters associated with N sample groups based on the first sample information and the first initial model of the sample objects on the first device; N is the number of object groups configured in the first initial model, and N is a positive integer greater than 1; the sample label of the first sample information is used to characterize the actual indicator parameters of the sample objects for the sample business data.

[0221] The number of sample objects is H, where H is a positive integer greater than or equal to N; the first sample information includes the first sample sub-information of each of the H sample objects on the first device;

[0222] The initial parameter acquisition module 100 includes: a grouping unit 1010, a sample group coding vector determination unit 1020, a sample group coding feature determination unit 1030, and an initial parameter determination unit 1040.

[0223] The grouping unit 1010 is used to obtain the number of object groups configured for the first initial model, and to group the H sample objects based on the number of object groups to obtain N sample groups; the sample group includes sample group j; j is a positive integer less than or equal to N;

[0224] The sample group encoding vector determination unit 1020 is used to obtain the first sample sub-information associated with each sample object in sample group j from H first sample sub-information, and to perform feature extraction processing on the obtained first sample sub-information to obtain the sample group encoding vector V corresponding to sample group j. j ;

[0225] The sample group coding feature determination unit 1030 is used to obtain the sample group coding features associated with the N sample groups when the sample group coding vector corresponding to each of the N sample groups is obtained.

[0226] The initial parameter determination unit 1040 is used to determine the initial model parameters associated with N sample groups based on the sample group coding features.

[0227] The specific implementation methods of the grouping unit 1010, the sample group coding vector determination unit 1020, the sample group coding feature determination unit 1030, and the initial parameter determination unit 1040 can be found in the above description. Figure 5 The description of step S205 in the corresponding embodiment will not be repeated here.

[0228] The sample group feature determination module 200 is used to perform self-attention feature extraction processing on the sample platform features, sample group encoding features and initial model parameters of the sample object in the first initial model to obtain the sample group attention features of the sample object for N sample groups; the sample platform features are obtained by the second device after extracting features from the second sample information of the sample object on the second device based on the second initial model.

[0229] The prediction module 300 is used to predict the prediction index parameters of the sample object for the sample business data based on the sample data features, sample platform features, and sample group attention features of the sample business data in the first initial model.

[0230] The loss determination module 400 is used to determine the total model loss corresponding to the first initial model based on the predicted index parameters, the actual index parameters, the sample group encoding features, and the sample group attention features; the total model loss is used to instruct the second device to perform federated learning training on the second initial model to obtain the second business model.

[0231] The loss determination module 400 includes: a first loss determination unit 4010, a second loss determination unit 4020, and a total loss determination unit 4030.

[0232] The first loss determination unit 4010 is used to determine the first model loss of the first initial model based on the sample group encoding features and the sample group attention features.

[0233] The second loss determination unit 4020 is used to determine the second model loss of the second initial model based on the predicted index parameters and the actual index parameters.

[0234] The total loss determination unit 4030 is used to superimpose the first model loss and the second model loss to obtain the total model loss corresponding to the first initial model.

[0235] The specific implementation methods of the first loss determination unit 4010, the second loss determination unit 4020, and the total loss determination unit 4030 can be found in the above description. Figure 5 The description of step S208 in the corresponding embodiment will not be repeated here.

[0236] The training module 500 is used to perform federated learning training on the first initial model based on the total model loss and the attention features of the sample group to obtain the first business model; the first business model is used to predict the business indicator parameters of the business object.

[0237] The training module 500 includes a training result determination unit 5010, a parameter adjustment unit 5020, and a business model determination unit 5030.

[0238] The training result determines unit 5010, which is used to perform federated learning training on the first initial model based on the total model loss to obtain the model training result;

[0239] The parameter adjustment unit 5020 is used to adjust the parameters of the initial model based on the attention features of the sample group and the total loss of the model if the model training result indicates that the first initial model after training does not meet the model convergence condition associated with the first initial model.

[0240] The business model determination unit 5030 is used to take the adjusted first initial model as a transition model, perform federated learning training on the transition model, and when the transition model after federated learning training meets the model convergence condition, the transition model that meets the model convergence condition is taken as the first business model.

[0241] The specific implementation methods of the training result determination unit 5010, parameter adjustment unit 5020, and business model determination unit 5030 can be found above. Figure 5 The description of step S209 in the corresponding embodiment will not be repeated here.

[0242] The second identifier set acquisition module 600 is used to acquire a second object identifier set associated with the second device based on the sample alignment request when a sample alignment request is received from the second device; the object identifiers in the second object identifier set are the object identifiers of historical objects recorded on the second device.

[0243] The second identifier set acquisition module 600 includes: an encrypted information acquisition unit 6010, a decryption processing unit 6020, a signature verification processing unit 6030, and an identifier set determination unit 6040.

[0244] The encrypted information acquisition unit 6010 is used to acquire the identifier encrypted information in the sample alignment request when the sample alignment request sent by the second device is received; the identifier encrypted information is obtained by the second device encrypting the identifier signature information and the platform object identifier set based on the first public key of the first device; the identifier signature information is obtained by the second device signing the platform object identifier set based on the second private key of the second device.

[0245] The decryption processing unit 6020 is used to decrypt the identifier encryption information based on the first private key of the first device to obtain the identifier signature information and the platform object identifier set.

[0246] The signature verification processing unit 6030 is used to obtain the second public key of the second device, and perform signature verification processing on the identifier signature information based on the second public key to obtain the signature verification result;

[0247] The identifier set determination unit 6040 is used to use the platform object identifier set as the second object identifier set associated with the second device when the verification result indicates that the verification is successful.

[0248] The specific implementation methods of the encrypted information acquisition unit 6010, decryption processing unit 6020, signature verification processing unit 6030, and identifier set determination unit 6040 can be found in the above description. Figure 6 The description of the second object identifier set in the corresponding embodiments will not be repeated here.

[0249] The first identifier set acquisition module 700 is used to acquire a first object identifier set associated with the first device based on the object identifiers of historical objects recorded on the first device.

[0250] The alignment processing module 800 is used to align the first object identifier set and the second object identifier set to obtain an identifier intersection set. The historical objects corresponding to the identifier objects in the identifier intersection set are used as sample objects that have an intersection relationship with the second device. All identifier objects in the identifier intersection set belong to the first object identifier set and the second object identifier set.

[0251] The identifier intersection return module 900 is used to return the identifier intersection set to the second device so that the second device can determine the sample platform characteristics of the sample objects in the identifier intersection set.

[0252] The specific implementation methods of the initial parameter acquisition module 100, sample group feature determination module 200, prediction module 300, loss determination module 400, training module 500, second identifier set acquisition module 600, first identifier set acquisition module 700, alignment processing module 800, and identifier intersection return module 900 can be found above. Figure 5 The descriptions of steps S201-S209 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0253] Further, please see Figure 11 , Figure 11 This is a schematic diagram of a computer device provided in an embodiment of this application. Figure 11 As shown, the computer device 1000 can be a first device, which is a computer device deployed with a first business model, and the first device can be the aforementioned Figure 1 The server 11F in the corresponding embodiment can also be the one described above. Figure 1The terminal device 1000 in the corresponding embodiment is any one of the terminal devices running the first service model in the terminal device cluster. For example, terminal device 100a. The computer device 1000 may include: at least one processor 1001, such as a CPU; at least one network interface 1004; a memory 1005; and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 11 As shown, the memory 1005, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a device control application program. In some embodiments, the computer device may further include... Figure 11 The user interface 1003 shown, for example, if the computer device is Figure 1 The terminal device (e.g., terminal device 100a) deployed with the first business model shown may also include the user interface 1003, wherein the user interface 1003 may include a display screen, a keyboard, etc.

[0254] exist Figure 11 In the computer device 1000 shown, the network interface 1004 is mainly used for network communication; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005.

[0255] It should be understood that the computer device 1000 described in the embodiments of this application can execute the foregoing text. Figure 3 and Figure 5 The description of the data processing method in the corresponding embodiment can also be performed as described above. Figure 9 The corresponding embodiment refers to the data processing device 1 or Figure 10 The description of the data processing device 2 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated here.

[0256] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium storing a computer program executed by the aforementioned data processing device 1 or data processing device 2, and the computer program includes program instructions. When the processor executes the program instructions, it can execute the aforementioned... Figure 3 or Figure 5 The description of the data processing method in the corresponding embodiments is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments related to this application, please refer to the description of the method embodiments of this application. As an example, program instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network. These multiple computing devices distributed across multiple locations and interconnected via a communication network can constitute a blockchain system.

[0257] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which are implemented when executed by a processor. Figure 3 and Figure 5 For details on the data processing methods provided in each step, please refer to [link / reference]. Figure 3 as well as Figure 5 The implementation methods provided for each step will not be elaborated here.

[0258] For further details, please see Figure 12 , Figure 12 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application. The data processing system 3 may include a data processing device 1a and a data processing device 2a. The data processing device 1a may be the one described above. Figure 9 Regarding the data processing device 1 in the corresponding embodiment, it is understood that the data processing device 1a can be integrated into the aforementioned first device, which is a computer device deployed with the first business model. The first device can be the aforementioned... Figure 1 The server 11F in the corresponding embodiment can also be the one described above. Figure 1 The terminal device in the terminal device cluster corresponding to the embodiment that runs the first service model, for example, terminal device 100a, will not be described in detail here. The data processing device 2a can be the one described above. Figure 10 The data processing device 2 in the corresponding embodiment can be understood to be integrated into the first device described above; therefore, it will not be described again here. Furthermore, the beneficial effects of using the same method will also not be described again. For technical details not disclosed in the embodiments of the data processing system involved in this application, please refer to the description of the method embodiments of this application.

[0259] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0260] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, The method is performed by a first device and includes: Obtain the object platform characteristics of the business object; the object platform characteristics are obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model; Based on the first business model, business model parameters associated with N object groups and object group encoding features associated with the N object groups are obtained; N is a positive integer greater than 1; the first business model and the second business model are obtained by federated learning of sample objects; In the first business model, the business model parameters, the object group encoding features, and the object platform features are subjected to self-attention feature extraction processing to obtain the object group attention features of the business object for the N object groups; In the first business model, based on the candidate data features of the candidate business data, the object platform features, and the object group attention features, the business indicator parameters of the business object for the candidate business data are predicted.

2. The method according to claim 1, characterized in that, The platform feature of the object is obtained by the first device when it receives a traffic request for the service object; the traffic request for the service object belongs to X traffic requests sent by the second device; X is a positive integer; The X traffic requests are selected by the second device when generating traffic requests corresponding to Y initial objects, based on the initial platform features corresponding to the Y initial objects and the third business model deployed on the second device; Y is a positive integer greater than or equal to X; the initial platform feature of an initial object is obtained by the second device after extracting features from the object information of the initial object on the second device based on the second business model.

3. The method according to claim 1, characterized in that, The business model parameters include the object group pooling vectors corresponding to the N object groups respectively; the object group encoding features include the object group encoding vectors corresponding to the N object groups respectively; the N object groups include object group i; i is a positive integer less than or equal to N; The step of performing self-attention feature extraction processing on the business model parameters, the object group encoding features, and the object platform features in the first business model to obtain the object group attention features of the business object for the N object groups includes: The object platform features are input into the self-attention layer in the first business model; Obtain the object group pooling vector K of object group i from N object group pooling vectors. i Obtain the object group encoding vector V of object group i from N object group encoding vectors. i ; Through the self-attention layer, the object platform features and the object group pooling vector K are processed. i and the object group encoding vector V i After performing correlation processing, the attention coefficient W corresponding to the object group i is obtained. i ; When the attention coefficients corresponding to each of the N object groups are obtained, the N attention coefficients are summed by the self-attention layer to obtain the object group attention features of the business object for the N object groups.

4. The method according to claim 1, characterized in that, The step of predicting the business indicator parameters of the business object for the candidate business data based on the candidate data features, the object platform features, and the object group attention features in the first business model includes: The feature splicing layer in the first business model is used to splice the object platform features and the object group attention features to obtain the business splicing features corresponding to the business object. Obtain the candidate data features obtained after feature extraction of candidate business data; The business splicing features and the candidate data features are input into the multilayer perceptron in the first business model. The multilayer perceptron performs feature extraction processing on the business splicing features and the candidate data features to predict the business indicator parameters of the business object for the candidate business data.

5. The method according to claim 1, characterized in that, The number of candidate service data is M, where M is a positive integer; The method further includes: When the business indicator parameters of the business object for each of the M candidate business data are obtained based on the first business model, the M business indicator parameters are used as M decision parameters. The M decision parameters are returned to the second device so that the second device can select the service data to be sent to the service object from the M candidate service data; the service data to be displayed is the top P candidate service data obtained based on the data sorting result; the data sorting result is obtained by the second device sorting the M candidate service data based on the sorting parameters corresponding to the M candidate service data respectively; the sorting parameter of a candidate service data is determined based on the decision parameter of a candidate service data; P is a positive integer less than or equal to M.

6. A data processing method, characterized in that, The method is performed by a first device and includes: Based on the first sample information of the sample object on the first device and the first initial model, sample group encoding features associated with N sample groups and initial model parameters associated with the N sample groups are obtained; N is the number of object groups configured in the first initial model, and N is a positive integer greater than 1; the sample label of the first sample information is used to characterize the actual indicator parameters of the sample object for the sample business data. In the first initial model, self-attention feature extraction processing is performed on the sample platform features, sample group encoding features, and initial model parameters of the sample object to obtain the sample group attention features of the sample object for N sample groups; the sample platform features are obtained by the second device based on the second initial model after extracting features from the second sample information of the sample object on the second device; In the first initial model, based on the sample data features of the sample business data, the sample platform features, and the sample group attention features, the prediction index parameters of the sample object for the sample business data are predicted. Based on the predicted index parameters, the actual index parameters, the sample group encoding features, and the sample group attention features, the total model loss corresponding to the first initial model is determined; the total model loss is used to instruct the second device to perform federated learning training on the second initial model to obtain the second business model. Based on the total loss of the model and the attention features of the sample group, the first initial model is trained by federated learning to obtain the first business model; the first business model is used to predict the business indicator parameters of the business object.

7. The method according to claim 6, characterized in that, The method further includes: Upon receiving a sample alignment request sent by the second device, a second object identifier set associated with the second device is obtained based on the sample alignment request; the object identifiers in the second object identifier set are the object identifiers of historical objects recorded on the second device. Based on the object identifiers of historical objects recorded on the first device, obtain the first object identifier set associated with the first device; Alignment processing is performed on the first object identifier set and the second object identifier set to obtain an identifier intersection set. The historical objects corresponding to the identifier objects in the identifier intersection set are taken as sample objects that have an intersection relationship with the second device. All identifier objects in the identifier intersection set belong to the first object identifier set and the second object identifier set. The set of identifier intersections is returned to the second device so that the second device can determine the sample platform characteristics of the sample objects in the set of identifier intersections.

8. The method according to claim 7, characterized in that, Upon receiving a sample alignment request sent by the second device, the step of obtaining a second object identifier set associated with the second device based on the sample alignment request includes: Upon receiving a sample alignment request sent by the second device, the identifier encryption information in the sample alignment request is obtained; the identifier encryption information is obtained by the second device encrypting the identifier signature information and the platform object identifier set based on the first device's first public key; the identifier signature information is obtained by the second device signing the platform object identifier set based on the second device's second private key. Based on the first private key of the first device, the encrypted identifier information is decrypted to obtain the identifier signature information and the platform object identifier set; Obtain the second public key of the second device, and perform signature verification processing on the identifier signature information based on the second public key to obtain the signature verification result; When the verification result indicates successful verification, the platform object identifier set is used as the second object identifier set associated with the second device.

9. The method according to claim 6, characterized in that, The number of sample objects is H, where H is a positive integer greater than or equal to N; the first sample information includes the first sample sub-information of each of the H sample objects on the first device; The step of obtaining sample group encoding features associated with N sample groups and initial model parameters associated with the N sample groups based on the first sample information of the sample objects on the first device and the first initial model includes: Obtain the number of object groups configured for the first initial model, and group the H sample objects based on the number of object groups to obtain N sample groups; each sample group includes sample group j; j is a positive integer less than or equal to N; First sample sub-information associated with each sample object in sample group j is obtained from H first sample sub-information pieces. Feature extraction processing is performed on the obtained first sample sub-information to obtain the sample group encoding vector V corresponding to sample group j. j ; When the sample group encoding vector corresponding to each of the N sample groups is obtained, the sample group encoding features associated with the N sample groups are obtained. Based on the coding features of the sample groups, the initial model parameters associated with the N sample groups are determined.

10. The method according to claim 6, characterized in that, The step of determining the total model loss corresponding to the first initial model based on the predicted index parameters, the actual index parameters, the sample group encoding features, and the sample group attention features includes: Based on the encoding features of the sample group and the attention features of the sample group, the first model loss of the first initial model is determined; Based on the predicted index parameters and the actual index parameters, determine the second model loss of the second initial model; The first model loss and the second model loss are superimposed to obtain the total model loss corresponding to the first initial model.

11. The method according to claim 6, characterized in that, The first initial model is trained using federated learning based on the total loss of the model and the attention features of the sample group to obtain the first business model, including: Based on the total model loss, the first initial model is trained using federated learning to obtain the model training result; If the model training result indicates that the first initial model after training does not meet the model convergence condition associated with the first initial model, then the parameters of the initial model are adjusted based on the attention features of the sample group and the total loss of the model. The adjusted first initial model is used as a transition model. Federated learning training is performed on the transition model until the transition model after federated learning training meets the model convergence condition. The transition model that meets the model convergence condition is used as the first business model.

12. A data processing apparatus, characterized in that, include: The platform feature acquisition module is used to acquire the platform features of business objects. The object platform feature is obtained by the second device after extracting features from the object information of the business object on the second device based on the second business model; The business parameter acquisition module is used to acquire business model parameters associated with N object groups and object group encoding features associated with the N object groups based on the first business model; N is a positive integer greater than 1; the first business model and the second business model are obtained by federated learning of sample objects; The object group feature determination module is used to perform self-attention feature extraction processing on the business model parameters, the object group encoding features and the object platform features in the first business model to obtain the object group attention features of the business object for the N object groups; The indicator parameter prediction module is used to predict the business indicator parameters of the business object for the candidate business data based on the candidate data features of the candidate business data, the object platform features, and the object group attention features in the first business model.

13. A data processing apparatus, characterized in that, include: The initial parameter acquisition module is used to acquire sample group encoding features associated with N sample groups and initial model parameters associated with the N sample groups based on the first sample information of the sample objects on the first device and the first initial model; N is the number of object groups configured in the first initial model, and N is a positive integer greater than 1; the sample label of the first sample information is used to characterize the actual indicator parameters of the sample objects for the sample business data. The sample group feature determination module is used to perform self-attention feature extraction processing on the sample platform features, sample group encoding features and initial model parameters of the sample object in the first initial model to obtain the sample group attention features of the sample object for N sample groups; the sample platform features are obtained by the second device after extracting features from the second sample information of the sample object on the second device based on the second initial model. The prediction module is used to predict the prediction index parameters of the sample object for the sample business data based on the sample data features of the sample business data, the sample platform features, and the sample group attention features in the first initial model. The loss determination module is used to determine the total model loss corresponding to the first initial model based on the predicted indicator parameters, the actual indicator parameters, the sample group encoding features, and the sample group attention features; the total model loss is used to instruct the second device to perform federated learning training on the second initial model to obtain the second business model; The training module is used to perform federated learning training on the first initial model based on the total loss of the model and the attention features of the sample group to obtain a first business model; the first business model is used to predict the business indicator parameters of the business object.

14. A computer device, characterized in that, include: Processor, memory, and network interface; The processor is connected to the memory and the network interface, wherein the network interface is used to provide data communication functions, the memory is used to store computer programs, and the processor is used to invoke the computer programs to cause the computer device to perform the method according to any one of claims 1 to 11.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1 to 11.

16. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium, the computer program being adapted to be read and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1 to 11.

Images